[llvm-bugs] [Bug 42741] New: LegacyDivergenceAnalysis: sync dependence misses some divergent uses outside a loop

via llvm-bugs llvm-bugs at lists.llvm.org
Wed Jul 24 09:16:40 PDT 2019


https://bugs.llvm.org/show_bug.cgi?id=42741

            Bug ID: 42741
           Summary: LegacyDivergenceAnalysis: sync dependence misses some
                    divergent uses outside a loop
           Product: libraries
           Version: trunk
          Hardware: PC
                OS: Linux
            Status: NEW
          Severity: enhancement
          Priority: P
         Component: Global Analyses
          Assignee: unassignedbugs at nondot.org
          Reporter: jay.foad at gmail.com
                CC: llvm-bugs at lists.llvm.org

Created attachment 22281
  --> https://bugs.llvm.org/attachment.cgi?id=22281&action=edit
test case

See the attached test case,
test/Analysis/LegacyDivergenceAnalysis/AMDGPU/syncbug.ll.

%tmp7 is divergent because it's sync-dependent on the divergent loop exit
condition %tmp5, but LegacyDivergenceAnalysis does not mark it as divergent.

Quoting from LegacyDivergenceAnalysis.cpp:

  // Propagation rule 2: if a value defined in a loop is used outside, the user
  // is sync dependent on the condition of the loop exits that dominate the
  // user. For example,

OK.

  // The algorithm used here handles both natural and unstructured loops. 
Given
  // a branch TI, we first compute its influence region, the union of all
simple
  // paths from TI to its immediate post dominator (IPostDom). Then, we search
  // for all the values defined in the influence region but used outside. All
  // these users are sync dependent on TI.

This algorithm doesn't appear to work. The influence region includes all blocks
from bb2 to bb8, so %tmp7 is never marked divergent by this algorithm because
it is not outside the influence region.

This might be related to 37185.

-- 
You are receiving this mail because:
You are on the CC list for the bug.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-bugs/attachments/20190724/c12b7f00/attachment.html>


More information about the llvm-bugs mailing list